Support encoder-only models without KV-Cache (#21270)

Signed-off-by: Max de Bayser <maxdebayser@gmail.com>
Signed-off-by: Max de Bayser <mbayser@br.ibm.com>
Co-authored-by: Russell Bryant <rbryant@redhat.com>
This commit is contained in:
Maximilien de Bayser
2025-07-26 10:09:52 -03:00
committed by GitHub
parent f27fdfc3ed
commit 1cd6eaba54
17 changed files with 352 additions and 99 deletions

View File

@@ -22,10 +22,12 @@ REVISION_ROBERTA = os.environ.get("REVISION", "main")
@pytest.mark.skipif(current_platform.is_rocm(),
reason="Xformers backend is not supported on ROCm.")
def test_model_loading_with_params(vllm_runner):
def test_model_loading_with_params(vllm_runner, monkeypatch):
"""
Test parameter weight loading with tp>1.
"""
# to use apply_model
monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")
with vllm_runner(model_name=MODEL_NAME,
revision=REVISION,
dtype="float16",
@@ -61,10 +63,12 @@ def test_model_loading_with_params(vllm_runner):
@pytest.mark.skipif(current_platform.is_rocm(),
reason="Xformers backend is not supported on ROCm.")
def test_roberta_model_loading_with_params(vllm_runner):
def test_roberta_model_loading_with_params(vllm_runner, monkeypatch):
"""
Test parameter weight loading with tp>1.
"""
# to use apply_model
monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")
with vllm_runner(model_name=MODEL_NAME_ROBERTA,
revision=REVISION_ROBERTA,
dtype="float16",
@@ -101,10 +105,12 @@ def test_roberta_model_loading_with_params(vllm_runner):
@pytest.mark.skipif(current_platform.is_rocm(),
reason="Xformers backend is not supported on ROCm.")
def test_facebook_roberta_model_loading_with_params(vllm_runner):
def test_facebook_roberta_model_loading_with_params(vllm_runner, monkeypatch):
"""
Test loading roberta-base model with no lm_head.
"""
# to use apply_model
monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1")
model_name = "FacebookAI/roberta-base"
with vllm_runner(model_name=model_name,
dtype="float16",